YecoAI

Official Release: Yeco-PII v1.0 · The Most Powerful PII Model, Built by Us

278M-parameter neural model for PII detection and anonymization. Native Italian, 6 European languages, average F1 0.966.

Marco Nasi

Founder & CEO

Official Release: Yeco-PII v1.0 · A New State of the Art for PII Detection, Built In-House

What We Are Releasing

The Numbers Speak for Themselves

Multilingual F1 Comparison, 6 European Languages

Language Coverage Radar

5.4× Smaller. Still Stronger.

Contextual Intelligence · Unique on the Market

Built for Production, Not Just Benchmarks

Availability: Two Ways to Get Yeco-PII

Enterprises, available today via API

Public API availability will follow as the enterprise program scales

This Is Only the Beginning

Explore More

Today, September 10, 2026, we are officially releasing Yeco-PII v1.0, the most powerful PII detection and anonymization model on the market. Built entirely in-house, from research to training to engineering. This is not just a product launch: it is the proof that YecoAI can build category-leading AI with a fraction of the resources.

Yeco-PII v1.0 is a neural model that detects and anonymizes personal data (PII) in business documents. It understands personal data from context, including names and surnames it has never seen during training, without regex, manual rules, or validation formulas.

We benchmarked Yeco-PII on a mix of real and synthetic data inspired by real-world scenarios, never seen during training, with independent third-party annotations (Ai4Privacy source). Same test for every model. The result, under the historical v1.0 protocol (17-tag gold): average F1 0.966 against 0.776 for the OpenAI Privacy Filter and 0.726 for Rizzo-PII, with Yeco-PII ahead on all 6 languages.

Yeco-PII v1.0 (278M) vs. rizzo-pii (0.3B) vs. OpenAI Privacy Filter (1.5B)

F1 score coverage across Italian, English, French, German, Spanish, and Dutch

OpenAI's Privacy Filter carries 1.5 billion parameters, 5.4× ours, and requires triple the memory. The extra weight does not translate into capability on European languages. Why? Because detection is not a brute-force problem. Checksum networks (mod-97 for IBAN, Luhn for cards) only catch what a formula can validate.

Names, addresses, organizations, and dates have no checksum. There, real learning is required, and that is exactly where competitors lose 20-30 points. Yeco-PII is 100% neural: it learns from context.

Yeco-PII distinguishes a date of birth (to censor) from a regulatory reference date (to keep), and recognizes never-seen names thanks to context: "Mr.", "born in". It separates birth dates from legal dates with dedicated tags, defines censorship policies in JSON or natural language applied instantly, and covers Italian-specific entities (tax IDs, VAT numbers, cadastral data) that generic models simply do not see. None of the competing models offer this capability.

Enterprise companies can start testing Yeco-PII right now through our API, evaluating it on their own real documents before committing. When ready, the model is licensed with a one-time payment and delivered in standard HuggingFace format for direct integration into their pipeline, running on their own infrastructure, with no recurring costs.

The roadmap is already set: the enterprise evaluation of Yeco-PII is available today through our API program, opening the model to individual developers, professionals, and smaller teams. Enterprise teams can request API access through our contact page.

Yeco-PII v1.0 remains proprietary YecoAI software. Reuse, redistribution, and resale without contract are prohibited. Full technical documentation is reserved for licensed clients.

This Is Only the Beginning

We want to be clear about what this release means. YecoAI is a small Italian company. No venture capital, no research army of thousands. And yet we built a PII model that outperforms models many times its size, measured on independent benchmarks, on a mix of real and synthetic data, in 6 languages.

This release is scaled as only the beginning: the demonstration that YecoAI can build category-leading AI with a fraction of the resources. The same methodology, compact models with rigorous evaluation and real-world focus, is already being applied to the next challenges. If a 278M-parameter model built in-house can beat a 1.5B giant, imagine what comes next.

"We didn't set out to compete. We set out to win. Yeco-PII is the proof that a focused, research-driven team can out-engineer players with 5× the resources. This is only the beginning of what YecoAI can do."
  • Marco Nasi, Founder YecoAI

Broader public API availability will follow as the enterprise program scales.

YecoAI · [email protected] · Full technical documentation reserved for licensed clients.

  • 278M parameters: 1.1 GB, standard HuggingFace format, integrable anywhere
  • 24 protected data types: names, tax IDs, VAT numbers, IBAN, credit cards, emails, phones, birth dates, addresses, ZIP codes, provinces, organizations, amounts, document references, cadastral data, license plates, and more
  • Native Italian: not a translation afterthought, the language it was born in
  • Certified on 6 European languages, with particularly strong multilingual support for English
  • ~3,700 characters/second on CPU: a 100-page document in about a minute, no GPU needed
  • ~1 GB RAM at scale: runs on edge devices, VMs, and standard CPU instances
  • Standard HuggingFace format: drops into any existing pipeline

Related content